Improving importance estimation in pool-based batch active learning for approximate linear regression.

Area of Science:

• Machine Learning
• Statistical Learning Theory

Background:

• Pool-based batch active learning aims to minimize generalization error by selecting training inputs from a test pool.
• The P-ALICE method addresses model misspecification using importance weighting based on input densities.
• Existing P-ALICE importance estimation assumes a small number of training samples, limiting its practical application.

Purpose of the Study:

• To propose an alternative importance estimation scheme for pool-based batch active learning.
• To address the limitations of existing P-ALICE methods when dealing with a large number of training samples.
• To validate the proposed importance estimation scheme through numerical experiments.

Main Methods:

• Developed an alternative importance estimation scheme for pool-based batch active learning.
• Utilized inclusion probability for estimating the importance of training samples.
• Conducted numerical experiments to demonstrate the validity of the proposed method.

Main Results:

• The proposed inclusion probability-based importance estimation scheme is effective.
• The new scheme overcomes the limitations of P-ALICE in scenarios with a large number of training samples.
• Numerical experiments confirmed the validity and improved performance of the proposed method.

Conclusions:

• The proposed importance estimation method offers a practical improvement for pool-based batch active learning.
• This approach enhances the robustness of active learning methods when dealing with large datasets.
• The findings suggest a more reliable way to select training data for improved model generalization.